search

huanglianghua / siamrpn-pytorch

A clean PyTorch implementation of SiamRPN tracker, evaluated on 7 datasets.
189 stars 32 forks source link

How did you determine the details of the network parameters? #5

Closed fzh0917 closed 4 years ago

fzh0917 commented 5 years ago

In your code, the numbers of kernels of five convolutional layers are 192, 512, 768, 768, 512, respectively. How did you get these numbers? In the SiamRPN paper, the authors say that they use the AlexNet for extracting features of template branch and detection branch, but the numbers of convolutional kernels of the AlexNet isn't same as yours. So, I am confused. I am looking forward to your reply, thank you!

Singapore-mor commented 4 years ago

Beside, I have some other problems:

  1. how do get stride and padding numbers of each layer?
  2. How do you know if each layer of the network needs MaxPooling, ReLU and BatchNormalization?

I have the same problems as you, in the essay the last kernels shoule be 256, while in this code is 512

Singapore-mor commented 4 years ago

maybe you can find these parameters in pysot? I haven't seen it, maybe some details in pysot are more convincible?

zxfans

邮箱：zhfans@126.com |

Signature is customized by Netease Mail Master

On 10/20/2019 19:25, Zhihong Fu wrote:

Beside, I have some other problems:

how do get stride and padding numbers of each layer? How do you know if each layer of the network needs MaxPooling, ReLU and BatchNormalization?

I have the same problems as you, in the essay the last kernels shoule be 256, while in this code is 512

I think that these numbers are the author(@huanglianghua)'s experiences, which make the performance of this reproduction better.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.